intermediate 2 min answer

What does a layered bronze-silver-gold structure actually buy, and where does it become ceremony?

medallionlayersreprocessingqualitybookingtrade-off
Show the full answer Hide the answer

What it buys

1. Raw data retained, so transformation bugs are recoverable. The bronze layer keeps what arrived, unmodified. When a transformation is found to be wrong — and it will be — history can be reprocessed. Without it, the error is permanent.

2. A clear place for quality enforcement. The bronze-to-silver step is where types are enforced, ranges checked and referential integrity validated, with a defined behaviour on failure.

3. Separation of concerns. Silver is cleaned and conformed; gold is shaped for consumption. Changing a business definition touches gold; changing a source's format touches bronze-to-silver.

4. Reprocessing without re-fetching. Partners may not offer a replay, so the raw landing is frequently the only copy.

Where it becomes ceremony

When every dataset passes through all three layers regardless of need.

  • A dataset with one consumer and no cleaning required does not need three physical copies.
  • A dataset that is already clean and conformed at source gains nothing from a silver layer that copies it.
  • Small reference data does not justify a pipeline per layer.

Each unnecessary layer is storage, compute, latency and a pipeline to operate — and the copies drift, which creates the reconciliation problem the structure was meant to avoid.

The principle that keeps it useful

Layers are a logical separation, not a mandatory physical one. The valuable properties are raw retained and quality enforced at a defined point. Whether silver is a materialised table or a view depends on whether anything reads it and whether the transformation is expensive.

The failure to design against

Quality checks that halt everything. A per-source pipeline with a broken feed should hold that source at its last known-good state and alert — not block every other source's updates, and not publish partial data either.

Scope failure per source, which is the same isolation principle that governs any multi-partner ingestion.